I am looking at the gold result for an information structure regression test. This is an SVO language with clause-final focus:
section=info-str focus-pos=clause-final topic-first=on c-focus-pos=preverbal
For the sentence PN TV CN, the gold has the following 5 trees:
The double-NP is narrow focus. The S over VP is subject extraction. The VP at the left edge is the information structure special head-complement phrase. The final tree is the canonical SVO tree.
Here’s the MRS for these 5 trees:
Am I right in thinking that this is essentially just two different structures:
and the rest is spurious?
In particular, MRS #2–3 have x3 which is not connected to anything. MRS#5 has a weird double ICONS.
So, getting rid of those three MRS should probably be counted as improvement?